Descripció del projecte
This doctoral project aims at the development and investigation on sentiment analysis of content written by users of Brandchats’ clients in social media. Current industrial sentiment analysis systems provide a single value (usually, positive/negative) per document, and do not take into account the different aspects on which people express their opinions nor the differences in opinion expressions across domains and different contexts. This project aims at covering these two aspects:
– to specify the aspects on which opinions are expressed (including the detection of Named Entities in documents)
– to create a system that takes into account the variability of opinion expressions across domains
1. Domain and languages:
The project will focus on the Financial and Pharmaceutical domains, as they are both commercially interesting fields and Brandchats has a great amount of data available for experiments. On the other hand, the type of document will be mainly tweets from Twitter due to its vast volume and popularity. Spanish and English will be the languages we concentrate on as annotated data under the Financial Domain in Spanish is already available and both are resourceful languages.
2. FIRST YEAR planning:
– Data Pre-processing and Normalization
– Named Entity Recognition (NER)
– Opinion Entity and Aspect Recognition
– Data Annotation for Evaluation
In the first year, the general objective is to build a Named Entity Recognizer (NER), i.e., a system that can detect entities about which opinions are expressed starting from the Financial Domain, while preparing data for experiments and evaluations. The NER system is expected to detect Named Entities and later categorize them into classes (i.e. Brands, Locations, Person…), using techniques like Pattern Recognition, Machine Learning or a combination of both. Another task will be finding entities with opinions (i.e. Financial Product, Online Service…), and aspects of these entities, by using a general
Sentiment Lexicon to identify opinion words first and extract target entities through linguistic patterns such as syntactic parses. Such technique will play an important role in further construction of a Contextualized Sentiment Lexicon.
Other tasks for this year are the pre-processing and normalization of texts and the preparation of annotated data under different domains and languages for system evaluation. These tasks are important as Twitter, which contains a rich set of opinions and great volume for both business intelligence and research purposes, is usually noisy due to its informality in expression.
3. SECOND YEAR planning:
– Domain Adaptation for Opinion Entity Recognition
– Domain Adaptation for General Sentiment Lexicon
– Detection and Categorization of Opinion Holder
For the second year, the main goal is to adapt existing systems to new domains as the polarity of some sentiment words depends on the domain or aspect. Therefore, the Opinion Entity Recognizer will be expected to be capable of working in multiple domains (i.e. Financial Domain to Pharmaceutical Domain) and the General Sentiment Lexicon will be enhanced with new Domain/Aspect Dependent Words (i.e. expand existing vocabulary under Financial Domain and Pharmaceutical Domain).
To construct a Domain/Aspect Dependent Sentiment Lexicon, we plan to use bootstrapping methods which allow the expansion of the dictionary starting from a small amount of data. Another approach can be building a basic Machine Learning as a General Sentiment Classifier and adapt it to a Domain Specific Classifier through methods like ‘co-training’.
In Sentiment Analysis, the subtopic of Detection and Classification of Opinion Holder is an interesting field, yet less explored. Depending on what kind of data we obtain, structured data from websites or data generated in Social Media, and in case there was no strong indication about the profession of the Opinion Holder, we will need to explore ways to find useful clues in the timeline of the user (i.e. through posts containing URL and looking for information in its meta description).
4. THIRD YEAR planning:
– Multi Domain Aspect Based Lexicon Integration
– Context Aware Sentiment Analysis
– General Polarity Summarization
The main task in the third year will be focusing on the integration of a multiple domain aspect based sentiment lexicon which will be used for Context Aware Sentiment Analysis.
The system will also be expected to summarize polarities on aspect level to general level taking into account the importance of each aspect as a feature in order to find out a more generic polarity about a certain entity.
According to observation and experience, the context dependent ambiguity problem in Sentiment Analysis is generally caused by domain and aspect information co-occurring with the opinion word, thus the integrated Multi Domain Aspect Based Sentiment Lexicon will be able to distinguish the difference accordingly.
Furthermore, factors like the category of the opinion word (i.e. verb, adjective, adverb…) and its uncovered relations within the context can also be important. Therefore, for better capture of the context information, we can also use additional techniques like word embeddings which allow a semantic representation of words in vector space for a better mapping of word relations and context information to gain more complex features and further improve the classifier’s performance.
Note: All experiment achievements can be submitted to relevant conferences or journals in related research fields.
